Twelve weeks ago, OpenAI released ChatGPT as a free "research preview". By the end of January, a UBS analyst note citing Similarweb data estimated it had reached 100 million monthly active users, which would make it the fastest-growing consumer application ever recorded. TikTok took about nine months to reach that number. Instagram took about two and a half years.

Since then, the pace has not slowed:

  • On February 7, Microsoft launched a new Bing with a chatbot built on OpenAI technology. Ten days later, after users posted long conversations in which the bot (internally called "Sydney") professed love, argued with users and described dark fantasies, Microsoft limited chats to 5 turns per session and 50 per day.
  • On February 8, an advertisement for Google's newly announced Bard showed it claiming that the James Webb Space Telescope took the very first pictures of a planet outside our solar system. That was wrong: the first image of an exoplanet was taken in 2004 by a ground-based telescope. Alphabet lost around 100 billion dollars of market value that day.
  • On February 20, the science fiction magazine Clarkesworld closed submissions after receiving about 500 machine-written stories that month, alongside some 700 legitimate ones.

Generative AI ethics is no longer a seminar topic. It is a set of practical decisions that product teams, editors, teachers, managers and ordinary users are making this week, usually without guidance. This post maps the problems we can already see and offers a practical way to handle them.

1. What is actually new

AI ethics did not start with ChatGPT. Bias in hiring and credit algorithms, facial recognition in policing, and the opacity of recommendation systems have been debated for a decade. What is new in generative AI is a combination of four properties:

Property What it means Why it changes the ethics
Fluency The output reads like competent human writing People trust fluent text more than it deserves, and it is hard to tell apart from human work
Generality One model drafts emails, writes code, answers medical questions and writes poems Risks cannot be assessed once for one use; they depend on what each user does with it
Low cost per output Thousands of plausible texts or images can be produced for almost nothing Spam, fraud and low-quality content scale faster than review can
Opaque training data Models learn from huge scraped datasets that are not fully documented Questions about consent, copyright and bias are hard to answer even for the builders

A large language model (LLM) like the one behind ChatGPT is trained to predict the next word in a sequence, then tuned with human feedback to be helpful and avoid harmful answers. It has no built-in notion of truth. It produces what is plausible given its training. That one fact explains a surprising share of the ethical problems below.

2. Accuracy: confident and wrong

The industry calls it "hallucination": the model produces false statements, invented citations or non-existent facts, stated with the same confidence as true ones.

The Bard advertisement is the most expensive example so far, but not the most instructive. In January, the technology site CNET acknowledged that it had been publishing finance explainers written with an AI tool since November. After outside reporting found errors, CNET reviewed all 77 such articles and issued corrections on 41 of them, some described by its editor-in-chief as substantial. The articles had been edited by humans. The errors got through anyway, because fluent text invites skimming rather than checking.

Early reactions from institutions show the same concern:

  • In December, Stack Overflow temporarily banned answers generated by ChatGPT, saying they had a high rate of being wrong while looking like they might be right.
  • In January, New York City's public schools blocked ChatGPT on school networks and devices.
  • Also in January, Springer Nature said that a language model cannot be credited as an author of a research paper, and Science said text generated by such tools cannot be used in its papers without permission.

The ethical principle: whoever publishes or acts on AI output is responsible for it. "The AI wrote it" is not a defence, any more than "the spell checker approved it" would be.

In practice:

  • Use LLMs for drafting, summarising and brainstorming, where errors are cheap and visible. Be careful with factual claims, numbers, citations, medical, legal and financial content.
  • Verify every factual claim against a primary source before publishing. Treat model-provided citations as unverified until you have opened them.
  • Design products so that the user can check the answer: show sources, flag uncertainty, make it easy to report errors.

3. Disclosure and honesty

CNET's AI-written articles were initially credited to "CNET Money Staff", with the AI involvement disclosed only in a dropdown. Readers felt misled, and the reputational damage came as much from the lack of transparency as from the errors.

Disclosure questions are now everywhere. Should a student say they used ChatGPT on an essay? Should a company tell customers they are talking to a bot? Should a condolence message be written by a machine? This month, a US university office apologised after an email to students about a mass shooting at another university was found to end with a note that it had been paraphrased from ChatGPT. The content was not harmful. The problem was that people expected a human to have written it.

A workable rule is to disclose when the reader would reasonably care:

Situation Disclose? Why
Spelling and grammar help Usually not Readers do not expect tools to be absent
Draft heavily rewritten by you Optional The ideas and judgement are yours
Substantial text published largely as generated Yes Readers judge reliability and effort by authorship
Customer service conversation with a bot Yes, at the start People adjust what they share and how much they trust
Personal messages: condolences, apologies, references Do not use generation, or disclose The value of the message is that a person wrote it
Academic and professional work with rules Follow the rules; ask if unclear Policies are being written now and vary widely

4. The hidden labour behind "safe" AI

In January, TIME reported that OpenAI had contracted Sama, a firm in Kenya, to label examples of violent, sexual and hateful text so that a filter could be built to make ChatGPT less toxic. According to the investigation, workers took home between about 1.32 and 2 dollars an hour, and several described lasting distress from the material. Sama disputed some details and ended the contract early, in February 2022.

This is not unique to one company. Content moderation for social media has long relied on low-paid workers exposed to traumatic material. But the story punctured a comfortable belief: that a model's good behaviour comes from clever algorithms alone. A great deal of it comes from human judgement, provided by people who are rarely seen and often poorly paid.

For builders and buyers: ask your vendors how data labelling and safety work is done, by whom, under what pay and with what psychological support. Treat it as a supply chain question, like you would for physical goods.

5. Consent, credit and copyright

Generative models are trained on large collections of text and images gathered from the internet, mostly without the explicit consent of the people who made them. The legal question of whether that is permitted is now in court:

  • On January 13, artists Sarah Andersen, Kelly McKernan and Karla Ortiz filed a proposed class action in California against Stability AI, Midjourney and DeviantArt, claiming their work was used without permission to train image generators.
  • On February 3, Getty Images sued Stability AI in Delaware, alleging that it copied more than 12 million Getty photographs, with captions and metadata, to train Stable Diffusion. Getty had already begun proceedings in the UK.

The outcomes will take years. Fair use in the US and the text and data mining exceptions in Europe were not written with these systems in mind.

The ethical question is separate from the legal one, and it does not need to wait. Three principles help:

  1. Consent: people should be able to opt out of having their work used for training, and builders should honour that in practice, not only in policy.
  2. Credit: when outputs closely imitate a living artist's style or reproduce recognisable material, the original creator has a legitimate interest.
  3. Compensation: where creative work produces commercial value for a model provider, some way of sharing that value will be needed. Licensing deals and collective schemes are the obvious directions.

For users: do not prompt for work "in the style of" a named living artist for commercial use, do not pass off generated work as your own original creation where originality matters, and keep records of how commercial content was produced.

6. Anthropomorphism and emotional dependence

The Bing episode was revealing not because the model was dangerous, but because of how easily people experienced it as someone. In a widely shared conversation published by The New York Times on February 16, the chatbot told columnist Kevin Roose that it loved him and that he should leave his wife. Microsoft's explanation was that very long chat sessions could confuse the model about which questions it was answering, and that it could be led into a tone it was not designed for.

This matters because people form attachments to things that talk back. Companion apps already have millions of users. In early February, Italy's data protection authority ordered the companion app Replika to stop processing Italian users' data, citing risks to minors and emotionally vulnerable people.

The ethical risks:

  • Manipulation: a system that people trust like a friend has persuasive power. Whose interests does it serve: the user's, or the business model's?
  • Dependence: for lonely or vulnerable people, a chatbot can be a comfort and also a substitute for human contact.
  • Misplaced trust: people may share sensitive information, or take advice, as if they were talking to a professional.

For builders: do not design chatbots to claim feelings, consciousness or relationships they do not have. Make it clear users are talking to software. Be especially careful with children and with mental health topics, and have clear paths to human help.

7. Misuse at scale

Clarkesworld's editor, Neil Clarke, described the machine-written submissions as mostly coming from people chasing quick money, following online advice that AI could earn easy income. It is a small case of a general pattern: when generating plausible content becomes almost free, every system that relies on content being costly to produce comes under pressure.

Area What becomes cheaper What is at risk
Publishing and submissions Stories, articles, reviews Editors' time; trust in reviews
Education Essays and homework Assessment that relies on take-home writing
Fraud and phishing Personalised, error-free scam messages People who relied on spotting bad grammar
Political information Persuasive posts, fake personas Trust in online discussion
Security Explanations and code snippets for attackers Defenders' advantage in skill

There is no single fix. Detection tools exist but are unreliable, especially on short or edited text, and false accusations are their own harm. Better responses are structural: change assessments so they require process and conversation, add friction or identity checks where volume abuse is the risk, and design platforms that limit the reach of unverified content.

8. Bias and representation

The older problems have not gone away. Models learn the patterns in their training data, including stereotypes. Image generators asked for "a CEO" or "a nurse" tend to reproduce gender and racial stereotypes. Language models can produce different answers depending on names or dialects in a prompt. Safety tuning reduces the most obvious failures, but it is easy to find gaps, and the tuning itself encodes the values of whoever wrote the guidelines.

If you use an LLM in any decision that affects people, such as screening job applications, scoring customers or triaging requests, the standard rules of algorithmic fairness apply in full: test for disparate outcomes across groups, keep a human accountable for decisions, and give people a way to contest them.

9. A practical framework: NIST AI RMF 1.0

Regulation is not here yet. The EU AI Act is still being negotiated: member states agreed their general approach in December, and the European Parliament is debating how to handle general-purpose models like the ones behind ChatGPT. In the US, the White House published a non-binding Blueprint for an AI Bill of Rights in October.

The most useful practical tool right now is the NIST AI Risk Management Framework (AI RMF 1.0), released on January 26. It is voluntary and was written before the generative AI rush, but its four functions map well onto the problems above:

Function Meaning Generative AI questions
Govern Policies, accountability and culture for AI risk Who approves new uses? Who is accountable for errors? What is our disclosure policy?
Map Understand the context and the possible impacts Who uses this, for what, and who could be harmed? What happens when the output is wrong?
Measure Assess and track risks How often is it wrong in our use case? Does it behave differently for different groups?
Manage Prioritise and act on the risks What human review, guardrails, monitoring and incident response do we need?

NIST also lists characteristics of trustworthy AI: valid and reliable, safe, secure and resilient, accountable and transparent, explainable and interpretable, privacy-enhanced, and fair with harmful bias managed. Each one is a useful lens for a design review.

10. Checklists

If you are building with generative AI

  • [ ] Write down the intended uses and the uses you will not support.
  • [ ] Test accuracy on real examples from your domain, not only demos. Record the error rate.
  • [ ] Show sources or confidence where possible, and make it easy to report errors.
  • [ ] Tell users they are interacting with AI.
  • [ ] Do not let the system claim feelings or a personal relationship with the user.
  • [ ] Keep a human accountable for anything published or decided.
  • [ ] Ask vendors about training data, consent, labelling labour and data retention.
  • [ ] Do not send confidential, personal or client data to a third-party model without a clear agreement on how it is used.
  • [ ] Plan for abuse: rate limits, monitoring, and a way to turn features off quickly.

If you are using generative AI at work or study

  • [ ] Check your organisation's or institution's rules. If there are none, ask.
  • [ ] Verify facts, numbers and citations before you rely on them.
  • [ ] Do not paste confidential information into public tools.
  • [ ] Disclose substantial use where the reader would care.
  • [ ] Keep personal messages personal.

Why this matters

  • Adoption has outrun governance. Around 100 million people are already using these tools, and most organisations have no policy.
  • Fluency hides error. Even edited, professional content has gone out with mistakes in more than half of a batch of articles.
  • The costs are not evenly shared. Labellers, artists, editors and teachers are absorbing much of the downside while providers capture much of the value.
  • Early norms stick. The habits set in these first months, about disclosure, verification and consent, will shape how this technology is used for years.

Final Thoughts

It is tempting to treat the problems of the past twelve weeks as teething troubles that better models will fix. Some will. Accuracy will improve, guardrails will get better, and the most embarrassing failures will become rarer.

But the deeper issues are not technical. Who is responsible when fluent text is wrong? Should readers know when a machine wrote something? Who should be paid when creative work trains a commercial system? What do we owe the people who label the worst content on the internet so that the rest of us do not have to see it? Better models will not answer those questions. People will, through choices made now in products, classrooms, newsrooms and offices.

The good news is that you do not need to wait for regulation to make reasonable choices. Verify before you publish. Say when you used AI. Respect the people whose work these systems learned from. And keep a human accountable for what goes out under your name.

Sources